lustre/osd-zfs/osd_object.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osd-zfs/osd_object.c
  33  *
  34  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  35  * Author: Mike Pershin <tappro@whamcloud.com>
  36  * Author: Johann Lombardi <johann@whamcloud.com>
  37  */
  38
  39 #define DEBUG_SUBSYSTEM S_OSD
  40
  41 #include <lustre_ver.h>
  42 #include <libcfs/libcfs.h>
  43 #include <obd_support.h>
  44 #include <lustre_net.h>
  45 #include <obd.h>
  46 #include <obd_class.h>
  47 #include <lustre_disk.h>
  48 #include <lustre_fid.h>
  49
  50 #include "osd_internal.h"
  51
  52 #include <sys/dnode.h>
  53 #include <sys/dbuf.h>
  54 #include <sys/spa.h>
  55 #include <sys/stat.h>
  56 #include <sys/zap.h>
  57 #include <sys/spa_impl.h>
  58 #include <sys/zfs_znode.h>
  59 #include <sys/dmu_tx.h>
  60 #include <sys/dmu_objset.h>
  61 #include <sys/dsl_prop.h>
  62 #include <sys/sa_impl.h>
  63 #include <sys/txg.h>
  64
  65 char *osd_obj_tag = "osd_object";
  66
  67 static struct dt_object_operations osd_obj_ops;
  68 static struct lu_object_operations osd_lu_obj_ops;
  69 extern struct dt_body_operations osd_body_ops;
  70 static struct dt_object_operations osd_obj_otable_it_ops;
  71
  72 extern struct kmem_cache *osd_object_kmem;
  73
  74 static void
  75 osd_object_sa_fini(struct osd_object *obj)
  76 {
  77         if (obj->oo_sa_hdl) {
  78                 sa_handle_destroy(obj->oo_sa_hdl);
  79                 obj->oo_sa_hdl = NULL;
  80         }
  81 }
  82
  83 static int
  84 osd_object_sa_init(struct osd_object *obj, struct osd_device *o)
  85 {
  86         int rc;
  87
  88         LASSERT(obj->oo_sa_hdl == NULL);
  89         LASSERT(obj->oo_dn != NULL);
  90
  91         rc = osd_sa_handle_get(obj);
  92         if (rc)
  93                 return rc;
  94
  95         /* Cache the xattr object id, valid for the life of the object */
  96         rc = -sa_lookup(obj->oo_sa_hdl, SA_ZPL_XATTR(o), &obj->oo_xattr, 8);
  97         if (rc == -ENOENT) {
  98                 obj->oo_xattr = ZFS_NO_OBJECT;
  99                 rc = 0;
 100         } else if (rc) {
 101                 osd_object_sa_fini(obj);
 102         }
 103
 104         return rc;
 105 }
 106
 107 /*
 108  * Add object to list of dirty objects in tx handle.
 109  */
 110 static void
 111 osd_object_sa_dirty_add(struct osd_object *obj, struct osd_thandle *oh)
 112 {
 113         if (!list_empty(&obj->oo_sa_linkage))
 114                 return;
 115
 116         down(&oh->ot_sa_lock);
 117         write_lock(&obj->oo_attr_lock);
 118         if (likely(list_empty(&obj->oo_sa_linkage)))
 119                 list_add(&obj->oo_sa_linkage, &oh->ot_sa_list);
 120         write_unlock(&obj->oo_attr_lock);
 121         up(&oh->ot_sa_lock);
 122 }
 123
 124 /*
 125  * Release spill block dbuf hold for all dirty SAs.
 126  */
 127 void osd_object_sa_dirty_rele(struct osd_thandle *oh)
 128 {
 129         struct osd_object *obj;
 130
 131         down(&oh->ot_sa_lock);
 132         while (!list_empty(&oh->ot_sa_list)) {
 133                 obj = list_entry(oh->ot_sa_list.next,
 134                                  struct osd_object, oo_sa_linkage);
 135                 sa_spill_rele(obj->oo_sa_hdl);
 136                 write_lock(&obj->oo_attr_lock);
 137                 list_del_init(&obj->oo_sa_linkage);
 138                 write_unlock(&obj->oo_attr_lock);
 139         }
 140         up(&oh->ot_sa_lock);
 141 }
 142
 143 /*
 144  * Update the SA and add the object to the dirty list.
 145  */
 146 int osd_object_sa_update(struct osd_object *obj, sa_attr_type_t type,
 147                          void *buf, uint32_t buflen, struct osd_thandle *oh)
 148 {
 149         int rc;
 150
 151         LASSERT(obj->oo_sa_hdl != NULL);
 152         LASSERT(oh->ot_tx != NULL);
 153
 154         rc = -sa_update(obj->oo_sa_hdl, type, buf, buflen, oh->ot_tx);
 155         osd_object_sa_dirty_add(obj, oh);
 156
 157         return rc;
 158 }
 159
 160 /*
 161  * Bulk update the SA and add the object to the dirty list.
 162  */
 163 static int
 164 osd_object_sa_bulk_update(struct osd_object *obj, sa_bulk_attr_t *attrs,
 165                           int count, struct osd_thandle *oh)
 166 {
 167         int rc;
 168
 169         LASSERT(obj->oo_sa_hdl != NULL);
 170         LASSERT(oh->ot_tx != NULL);
 171
 172         rc = -sa_bulk_update(obj->oo_sa_hdl, attrs, count, oh->ot_tx);
 173         osd_object_sa_dirty_add(obj, oh);
 174
 175         return rc;
 176 }
 177
 178 /*
 179  * Retrieve the attributes of a DMU object
 180  */
 181 int __osd_object_attr_get(const struct lu_env *env, struct osd_device *o,
 182                           struct osd_object *obj, struct lu_attr *la)
 183 {
 184         struct osa_attr *osa = &osd_oti_get(env)->oti_osa;
 185         sa_bulk_attr_t  *bulk = osd_oti_get(env)->oti_attr_bulk;
 186         int              cnt = 0;
 187         int              rc;
 188         ENTRY;
 189
 190         LASSERT(obj->oo_dn != NULL);
 191
 192         la->la_valid |= LA_ATIME | LA_MTIME | LA_CTIME | LA_MODE | LA_TYPE |
 193                         LA_SIZE | LA_UID | LA_GID | LA_FLAGS | LA_NLINK;
 194
 195         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(o), NULL, osa->atime, 16);
 196         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(o), NULL, osa->mtime, 16);
 197         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CTIME(o), NULL, osa->ctime, 16);
 198         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(o), NULL, &osa->mode, 8);
 199         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_SIZE(o), NULL, &osa->size, 8);
 200         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_LINKS(o), NULL, &osa->nlink, 8);
 201         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_UID(o), NULL, &osa->uid, 8);
 202         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GID(o), NULL, &osa->gid, 8);
 203         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(o), NULL, &osa->flags, 8);
 204         LASSERT(cnt <= ARRAY_SIZE(osd_oti_get(env)->oti_attr_bulk));
 205
 206         rc = -sa_bulk_lookup(obj->oo_sa_hdl, bulk, cnt);
 207         if (rc)
 208                 GOTO(out_sa, rc);
 209
 210         la->la_atime = osa->atime[0];
 211         la->la_mtime = osa->mtime[0];
 212         la->la_ctime = osa->ctime[0];
 213         la->la_mode = osa->mode;
 214         la->la_uid = osa->uid;
 215         la->la_gid = osa->gid;
 216         la->la_nlink = osa->nlink;
 217         la->la_flags = attrs_zfs2fs(osa->flags);
 218         la->la_size = osa->size;
 219
 220         /* Try to get extra flag from LMA. Right now, only LMAI_ORPHAN
 221          * flags is stored in LMA, and it is only for orphan directory */
 222         if (S_ISDIR(la->la_mode) && dt_object_exists(&obj->oo_dt)) {
 223                 struct osd_thread_info *info = osd_oti_get(env);
 224                 struct lustre_mdt_attrs *lma;
 225                 struct lu_buf buf;
 226
 227                 lma = (struct lustre_mdt_attrs *)info->oti_buf;
 228                 buf.lb_buf = lma;
 229                 buf.lb_len = sizeof(info->oti_buf);
 230                 rc = osd_xattr_get(env, &obj->oo_dt, &buf, XATTR_NAME_LMA);
 231                 if (rc > 0) {
 232                         rc = 0;
 233                         lma->lma_incompat = le32_to_cpu(lma->lma_incompat);
 234                         obj->oo_lma_flags =
 235                                 lma_to_lustre_flags(lma->lma_incompat);
 236
 237                 } else if (rc == -ENODATA) {
 238                         rc = 0;
 239                 }
 240         }
 241
 242         if (S_ISCHR(la->la_mode) || S_ISBLK(la->la_mode)) {
 243                 rc = -sa_lookup(obj->oo_sa_hdl, SA_ZPL_RDEV(o), &osa->rdev, 8);
 244                 if (rc)
 245                         GOTO(out_sa, rc);
 246                 la->la_rdev = osa->rdev;
 247                 la->la_valid |= LA_RDEV;
 248         }
 249 out_sa:
 250
 251         RETURN(rc);
 252 }
 253
 254 int __osd_obj2dnode(const struct lu_env *env, objset_t *os,
 255                    uint64_t oid, dnode_t **dnp)
 256 {
 257         dmu_object_info_t *doi = &osd_oti_get(env)->oti_doi;
 258         dmu_buf_t *db;
 259         dmu_buf_impl_t *dbi;
 260         int rc;
 261
 262         rc = dmu_bonus_hold(os, oid, osd_obj_tag, &db);
 263         if (rc)
 264                 return rc;
 265
 266         dbi = (dmu_buf_impl_t *)db;
 267         DB_DNODE_ENTER(dbi);
 268         *dnp = DB_DNODE(dbi);
 269
 270         LASSERT(*dnp != NULL);
 271         dmu_object_info_from_dnode(*dnp, doi);
 272         if (unlikely (oid != DMU_USERUSED_OBJECT &&
 273             oid != DMU_GROUPUSED_OBJECT && doi->doi_bonus_type != DMU_OT_SA)) {
 274                 osd_dnode_rele(*dnp);
 275                 *dnp = NULL;
 276                 return -EINVAL;
 277         }
 278
 279         return 0;
 280 }
 281
 282 /*
 283  * Concurrency: no concurrent access is possible that early in object
 284  * life-cycle.
 285  */
 286 struct lu_object *osd_object_alloc(const struct lu_env *env,
 287                                    const struct lu_object_header *hdr,
 288                                    struct lu_device *d)
 289 {
 290         struct osd_object *mo;
 291
 292         OBD_SLAB_ALLOC_PTR_GFP(mo, osd_object_kmem, GFP_NOFS);
 293         if (mo != NULL) {
 294                 struct lu_object *l;
 295
 296                 l = &mo->oo_dt.do_lu;
 297                 dt_object_init(&mo->oo_dt, NULL, d);
 298                 mo->oo_dt.do_ops = &osd_obj_ops;
 299                 l->lo_ops = &osd_lu_obj_ops;
 300                 INIT_LIST_HEAD(&mo->oo_sa_linkage);
 301                 INIT_LIST_HEAD(&mo->oo_unlinked_linkage);
 302                 init_rwsem(&mo->oo_sem);
 303                 init_rwsem(&mo->oo_guard);
 304                 rwlock_init(&mo->oo_attr_lock);
 305                 mo->oo_destroy = OSD_DESTROY_NONE;
 306                 return l;
 307         } else {
 308                 return NULL;
 309         }
 310 }
 311
 312 /*
 313  * Concurrency: shouldn't matter.
 314  */
 315 int osd_object_init0(const struct lu_env *env, struct osd_object *obj)
 316 {
 317         struct osd_device       *osd = osd_obj2dev(obj);
 318         const struct lu_fid     *fid = lu_object_fid(&obj->oo_dt.do_lu);
 319         int                      rc = 0;
 320         ENTRY;
 321
 322         if (obj->oo_dn == NULL)
 323                 RETURN(0);
 324
 325         /* object exist */
 326
 327         rc = osd_object_sa_init(obj, osd);
 328         if (rc)
 329                 RETURN(rc);
 330
 331         /* cache attrs in object */
 332         rc = __osd_object_attr_get(env, osd, obj, &obj->oo_attr);
 333         if (rc)
 334                 RETURN(rc);
 335
 336         if (likely(!fid_is_acct(fid)))
 337                 /* no body operations for accounting objects */
 338                 obj->oo_dt.do_body_ops = &osd_body_ops;
 339
 340         /*
 341          * initialize object before marking it existing
 342          */
 343         obj->oo_dt.do_lu.lo_header->loh_attr |= obj->oo_attr.la_mode & S_IFMT;
 344
 345         smp_mb();
 346         obj->oo_dt.do_lu.lo_header->loh_attr |= LOHA_EXISTS;
 347
 348         RETURN(0);
 349 }
 350
 351 static int osd_check_lma(const struct lu_env *env, struct osd_object *obj)
 352 {
 353         struct osd_thread_info  *info = osd_oti_get(env);
 354         struct lu_buf           buf;
 355         int                     rc;
 356         struct lustre_mdt_attrs *lma;
 357         ENTRY;
 358
 359         CLASSERT(sizeof(info->oti_buf) >= sizeof(*lma));
 360         lma = (struct lustre_mdt_attrs *)info->oti_buf;
 361         buf.lb_buf = lma;
 362         buf.lb_len = sizeof(info->oti_buf);
 363
 364         rc = osd_xattr_get(env, &obj->oo_dt, &buf, XATTR_NAME_LMA);
 365         if (rc > 0) {
 366                 rc = 0;
 367                 lustre_lma_swab(lma);
 368                 if (unlikely((lma->lma_incompat & ~LMA_INCOMPAT_SUPP) ||
 369                              CFS_FAIL_CHECK(OBD_FAIL_OSD_LMA_INCOMPAT))) {
 370                         CWARN("%s: unsupported incompat LMA feature(s) %#x for "
 371                               "fid = "DFID"\n", osd_obj2dev(obj)->od_svname,
 372                               lma->lma_incompat & ~LMA_INCOMPAT_SUPP,
 373                               PFID(lu_object_fid(&obj->oo_dt.do_lu)));
 374                         rc = -EOPNOTSUPP;
 375                 }
 376         } else if (rc == -ENODATA) {
 377                 /* haven't initialize LMA xattr */
 378                 rc = 0;
 379         }
 380
 381         RETURN(rc);
 382 }
 383
 384 /*
 385  * Concurrency: no concurrent access is possible that early in object
 386  * life-cycle.
 387  */
 388 static int osd_object_init(const struct lu_env *env, struct lu_object *l,
 389                            const struct lu_object_conf *conf)
 390 {
 391         struct osd_object       *obj = osd_obj(l);
 392         struct osd_device       *osd = osd_obj2dev(obj);
 393         uint64_t                 oid;
 394         int                      rc;
 395         ENTRY;
 396
 397         LASSERT(osd_invariant(obj));
 398
 399         if (fid_is_otable_it(&l->lo_header->loh_fid)) {
 400                 obj->oo_dt.do_ops = &osd_obj_otable_it_ops;
 401                 l->lo_header->loh_attr |= LOHA_EXISTS;
 402                 RETURN(0);
 403         }
 404
 405         if (conf != NULL && conf->loc_flags & LOC_F_NEW)
 406                 GOTO(out, rc = 0);
 407
 408         rc = osd_fid_lookup(env, osd, lu_object_fid(l), &oid);
 409         if (rc == 0) {
 410                 LASSERT(obj->oo_dn == NULL);
 411                 rc = __osd_obj2dnode(env, osd->od_os, oid, &obj->oo_dn);
 412                 /* EEXIST will be returned if object is being deleted in ZFS */
 413                 if (rc == -EEXIST) {
 414                         rc = 0;
 415                         GOTO(out, rc);
 416                 }
 417                 if (rc != 0) {
 418                         CERROR("%s: lookup "DFID"/%#llx failed: rc = %d\n",
 419                                osd->od_svname, PFID(lu_object_fid(l)), oid, rc);
 420                         GOTO(out, rc);
 421                 }
 422                 LASSERT(obj->oo_dn);
 423                 rc = osd_object_init0(env, obj);
 424                 if (rc != 0)
 425                         GOTO(out, rc);
 426
 427                 rc = osd_check_lma(env, obj);
 428                 if (rc != 0)
 429                         GOTO(out, rc);
 430         } else if (rc == -ENOENT) {
 431                 rc = 0;
 432         }
 433         LASSERT(osd_invariant(obj));
 434 out:
 435         RETURN(rc);
 436 }
 437
 438 /*
 439  * Concurrency: no concurrent access is possible that late in object
 440  * life-cycle.
 441  */
 442 static void osd_object_free(const struct lu_env *env, struct lu_object *l)
 443 {
 444         struct osd_object *obj = osd_obj(l);
 445
 446         LASSERT(osd_invariant(obj));
 447
 448         dt_object_fini(&obj->oo_dt);
 449         OBD_SLAB_FREE_PTR(obj, osd_object_kmem);
 450 }
 451
 452 static int
 453 osd_object_unlinked_add(struct osd_object *obj, struct osd_thandle *oh)
 454 {
 455         int rc = -EBUSY;
 456
 457         LASSERT(obj->oo_destroy == OSD_DESTROY_ASYNC);
 458
 459         /* the object is supposed to be exclusively locked by
 460          * the caller (osd_object_destroy()), while the transaction
 461          * (oh) is per-thread and not shared */
 462         if (likely(list_empty(&obj->oo_unlinked_linkage))) {
 463                 list_add(&obj->oo_unlinked_linkage, &oh->ot_unlinked_list);
 464                 rc = 0;
 465         }
 466
 467         return rc;
 468 }
 469
 470 /* Default to max data size covered by a level-1 indirect block */
 471 static unsigned long osd_sync_destroy_max_size =
 472         1UL << (DN_MAX_INDBLKSHIFT - SPA_BLKPTRSHIFT + SPA_MAXBLOCKSHIFT);
 473 module_param(osd_sync_destroy_max_size, ulong, 0444);
 474 MODULE_PARM_DESC(osd_sync_destroy_max_size, "Maximum object size to use synchronous destroy.");
 475
 476 static inline void
 477 osd_object_set_destroy_type(struct osd_object *obj)
 478 {
 479         /*
 480          * Lock-less OST_WRITE can race with OST_DESTROY, so set destroy type
 481          * only once and use it consistently thereafter.
 482          */
 483         down_write(&obj->oo_guard);
 484         if (obj->oo_destroy == OSD_DESTROY_NONE) {
 485                 if (obj->oo_attr.la_size <= osd_sync_destroy_max_size)
 486                         obj->oo_destroy = OSD_DESTROY_SYNC;
 487                 else /* Larger objects are destroyed asynchronously */
 488                         obj->oo_destroy = OSD_DESTROY_ASYNC;
 489         }
 490         up_write(&obj->oo_guard);
 491 }
 492
 493 static int osd_declare_object_destroy(const struct lu_env *env,
 494                                       struct dt_object *dt,
 495                                       struct thandle *th)
 496 {
 497         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
 498         struct osd_object       *obj = osd_dt_obj(dt);
 499         struct osd_device       *osd = osd_obj2dev(obj);
 500         struct osd_thandle      *oh;
 501         int                      rc;
 502         uint64_t                 zapid;
 503         ENTRY;
 504
 505         LASSERT(th != NULL);
 506         LASSERT(dt_object_exists(dt));
 507
 508         oh = container_of0(th, struct osd_thandle, ot_super);
 509         LASSERT(oh->ot_tx != NULL);
 510
 511         /* declare that we'll remove object from fid-dnode mapping */
 512         zapid = osd_get_name_n_idx(env, osd, fid, NULL, 0);
 513         dmu_tx_hold_zap(oh->ot_tx, zapid, FALSE, NULL);
 514
 515         osd_declare_xattrs_destroy(env, obj, oh);
 516
 517         /* one less inode */
 518         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 519                                obj->oo_attr.la_gid, -1, oh, false, NULL, false);
 520         if (rc)
 521                 RETURN(rc);
 522
 523         /* data to be truncated */
 524         rc = osd_declare_quota(env, osd, obj->oo_attr.la_uid,
 525                                obj->oo_attr.la_gid, 0, oh, true, NULL, false);
 526         if (rc)
 527                 RETURN(rc);
 528
 529         osd_object_set_destroy_type(obj);
 530         if (obj->oo_destroy == OSD_DESTROY_SYNC)
 531                 dmu_tx_hold_free(oh->ot_tx, obj->oo_dn->dn_object,
 532                                  0, DMU_OBJECT_END);
 533         else
 534                 dmu_tx_hold_zap(oh->ot_tx, osd->od_unlinkedid, TRUE, NULL);
 535
 536         /* will help to find FID->ino when this object is being
 537          * added to PENDING/ */
 538         osd_idc_find_and_init(env, osd, obj);
 539
 540         RETURN(0);
 541 }
 542
 543 static int osd_object_destroy(const struct lu_env *env,
 544                               struct dt_object *dt, struct thandle *th)
 545 {
 546         struct osd_thread_info  *info = osd_oti_get(env);
 547         char                    *buf = info->oti_str;
 548         struct osd_object       *obj = osd_dt_obj(dt);
 549         struct osd_device       *osd = osd_obj2dev(obj);
 550         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
 551         struct osd_thandle      *oh;
 552         int                      rc;
 553         uint64_t                 oid, zapid;
 554         ENTRY;
 555
 556         down_write(&obj->oo_guard);
 557
 558         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 559                 GOTO(out, rc = -ENOENT);
 560
 561         LASSERT(obj->oo_dn != NULL);
 562
 563         oh = container_of0(th, struct osd_thandle, ot_super);
 564         LASSERT(oh != NULL);
 565         LASSERT(oh->ot_tx != NULL);
 566
 567         /* remove obj ref from index dir (it depends) */
 568         zapid = osd_get_name_n_idx(env, osd, fid, buf, sizeof(info->oti_str));
 569         rc = -zap_remove(osd->od_os, zapid, buf, oh->ot_tx);
 570         if (rc) {
 571                 CERROR("%s: zap_remove(%s) failed: rc = %d\n",
 572                        osd->od_svname, buf, rc);
 573                 GOTO(out, rc);
 574         }
 575
 576         rc = osd_xattrs_destroy(env, obj, oh);
 577         if (rc) {
 578                 CERROR("%s: cannot destroy xattrs for %s: rc = %d\n",
 579                        osd->od_svname, buf, rc);
 580                 GOTO(out, rc);
 581         }
 582
 583         oid = obj->oo_dn->dn_object;
 584         if (unlikely(obj->oo_destroy == OSD_DESTROY_NONE)) {
 585                 /* this may happen if the destroy wasn't declared
 586                  * e.g. when the object is created and then destroyed
 587                  * in the same transaction - we don't need additional
 588                  * space for destroy specifically */
 589                 LASSERT(obj->oo_attr.la_size <= osd_sync_destroy_max_size);
 590                 rc = -dmu_object_free(osd->od_os, oid, oh->ot_tx);
 591                 if (rc)
 592                         CERROR("%s: failed to free %s %llu: rc = %d\n",
 593                                osd->od_svname, buf, oid, rc);
 594         } else if (obj->oo_destroy == OSD_DESTROY_SYNC) {
 595                 rc = -dmu_object_free(osd->od_os, oid, oh->ot_tx);
 596                 if (rc)
 597                         CERROR("%s: failed to free %s %llu: rc = %d\n",
 598                                osd->od_svname, buf, oid, rc);
 599         } else { /* asynchronous destroy */
 600                 rc = osd_object_unlinked_add(obj, oh);
 601                 if (rc)
 602                         GOTO(out, rc);
 603
 604                 rc = -zap_add_int(osd->od_os, osd->od_unlinkedid,
 605                                   oid, oh->ot_tx);
 606                 if (rc)
 607                         CERROR("%s: zap_add_int() failed %s %llu: rc = %d\n",
 608                                osd->od_svname, buf, oid, rc);
 609         }
 610
 611 out:
 612         /* not needed in the cache anymore */
 613         set_bit(LU_OBJECT_HEARD_BANSHEE, &dt->do_lu.lo_header->loh_flags);
 614         if (rc == 0)
 615                 obj->oo_destroyed = 1;
 616         up_write(&obj->oo_guard);
 617         RETURN (0);
 618 }
 619
 620 static void osd_object_delete(const struct lu_env *env, struct lu_object *l)
 621 {
 622         struct osd_object *obj = osd_obj(l);
 623
 624         if (obj->oo_dn != NULL) {
 625                 osd_object_sa_fini(obj);
 626                 if (obj->oo_sa_xattr) {
 627                         nvlist_free(obj->oo_sa_xattr);
 628                         obj->oo_sa_xattr = NULL;
 629                 }
 630                 osd_dnode_rele(obj->oo_dn);
 631                 list_del(&obj->oo_sa_linkage);
 632                 obj->oo_dn = NULL;
 633         }
 634 }
 635
 636 /*
 637  * Concurrency: ->loo_object_release() is called under site spin-lock.
 638  */
 639 static void osd_object_release(const struct lu_env *env,
 640                                struct lu_object *l)
 641 {
 642 }
 643
 644 /*
 645  * Concurrency: shouldn't matter.
 646  */
 647 static int osd_object_print(const struct lu_env *env, void *cookie,
 648                             lu_printer_t p, const struct lu_object *l)
 649 {
 650         struct osd_object *o = osd_obj(l);
 651
 652         return (*p)(env, cookie, LUSTRE_OSD_ZFS_NAME"-object@%p", o);
 653 }
 654
 655 static void osd_object_read_lock(const struct lu_env *env,
 656                                  struct dt_object *dt, unsigned role)
 657 {
 658         struct osd_object *obj = osd_dt_obj(dt);
 659
 660         LASSERT(osd_invariant(obj));
 661
 662         down_read_nested(&obj->oo_sem, role);
 663 }
 664
 665 static void osd_object_write_lock(const struct lu_env *env,
 666                                   struct dt_object *dt, unsigned role)
 667 {
 668         struct osd_object *obj = osd_dt_obj(dt);
 669
 670         LASSERT(osd_invariant(obj));
 671
 672         down_write_nested(&obj->oo_sem, role);
 673 }
 674
 675 static void osd_object_read_unlock(const struct lu_env *env,
 676                                    struct dt_object *dt)
 677 {
 678         struct osd_object *obj = osd_dt_obj(dt);
 679
 680         LASSERT(osd_invariant(obj));
 681         up_read(&obj->oo_sem);
 682 }
 683
 684 static void osd_object_write_unlock(const struct lu_env *env,
 685                                     struct dt_object *dt)
 686 {
 687         struct osd_object *obj = osd_dt_obj(dt);
 688
 689         LASSERT(osd_invariant(obj));
 690         up_write(&obj->oo_sem);
 691 }
 692
 693 static int osd_object_write_locked(const struct lu_env *env,
 694                                    struct dt_object *dt)
 695 {
 696         struct osd_object *obj = osd_dt_obj(dt);
 697         int rc = 1;
 698
 699         LASSERT(osd_invariant(obj));
 700
 701         if (down_write_trylock(&obj->oo_sem)) {
 702                 rc = 0;
 703                 up_write(&obj->oo_sem);
 704         }
 705         return rc;
 706 }
 707
 708 static int osd_attr_get(const struct lu_env *env,
 709                         struct dt_object *dt,
 710                         struct lu_attr *attr)
 711 {
 712         struct osd_object       *obj = osd_dt_obj(dt);
 713         uint64_t                 blocks;
 714         uint32_t                 blksize;
 715         int                      rc = 0;
 716
 717         down_read(&obj->oo_guard);
 718
 719         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 720                 GOTO(out, rc = -ENOENT);
 721
 722         LASSERT(osd_invariant(obj));
 723         LASSERT(obj->oo_dn);
 724
 725         read_lock(&obj->oo_attr_lock);
 726         *attr = obj->oo_attr;
 727         if (obj->oo_lma_flags & LUSTRE_ORPHAN_FL)
 728                 attr->la_flags |= LUSTRE_ORPHAN_FL;
 729         read_unlock(&obj->oo_attr_lock);
 730
 731         /* with ZFS_DEBUG zrl_add_debug() called by DB_DNODE_ENTER()
 732          * from within sa_object_size() can block on a mutex, so
 733          * we can't call sa_object_size() holding rwlock */
 734         sa_object_size(obj->oo_sa_hdl, &blksize, &blocks);
 735         /* we do not control size of indices, so always calculate
 736          * it from number of blocks reported by DMU */
 737         if (S_ISDIR(attr->la_mode))
 738                 attr->la_size = 512 * blocks;
 739         /* Block size may be not set; suggest maximal I/O transfers. */
 740         if (blksize == 0)
 741                 blksize = osd_spa_maxblocksize(
 742                         dmu_objset_spa(osd_obj2dev(obj)->od_os));
 743
 744         attr->la_blksize = blksize;
 745         attr->la_blocks = blocks;
 746         attr->la_valid |= LA_BLOCKS | LA_BLKSIZE;
 747
 748 out:
 749         up_read(&obj->oo_guard);
 750         return rc;
 751 }
 752
 753 /* Simple wrapper on top of qsd API which implement quota transfer for osd
 754  * setattr needs. As a reminder, only the root user can change ownership of
 755  * a file, that's why EDQUOT & EINPROGRESS errors are discarded */
 756 static inline int qsd_transfer(const struct lu_env *env,
 757                                struct qsd_instance *qsd,
 758                                struct lquota_trans *trans, int qtype,
 759                                __u64 orig_id, __u64 new_id, __u64 bspace,
 760                                struct lquota_id_info *qi)
 761 {
 762         int     rc;
 763
 764         if (unlikely(qsd == NULL))
 765                 return 0;
 766
 767         LASSERT(qtype >= 0 && qtype < LL_MAXQUOTAS);
 768         qi->lqi_type = qtype;
 769
 770         /* inode accounting */
 771         qi->lqi_is_blk = false;
 772
 773         /* one more inode for the new owner ... */
 774         qi->lqi_id.qid_uid = new_id;
 775         qi->lqi_space      = 1;
 776         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 777         if (rc == -EDQUOT || rc == -EINPROGRESS)
 778                 rc = 0;
 779         if (rc)
 780                 return rc;
 781
 782         /* and one less inode for the current id */
 783         qi->lqi_id.qid_uid = orig_id;;
 784         qi->lqi_space      = -1;
 785         /* can't get EDQUOT when reducing usage */
 786         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 787         if (rc == -EINPROGRESS)
 788                 rc = 0;
 789         if (rc)
 790                 return rc;
 791
 792         /* block accounting */
 793         qi->lqi_is_blk = true;
 794
 795         /* more blocks for the new owner ... */
 796         qi->lqi_id.qid_uid = new_id;
 797         qi->lqi_space      = bspace;
 798         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 799         if (rc == -EDQUOT || rc == -EINPROGRESS)
 800                 rc = 0;
 801         if (rc)
 802                 return rc;
 803
 804         /* and finally less blocks for the current owner */
 805         qi->lqi_id.qid_uid = orig_id;
 806         qi->lqi_space      = -bspace;
 807         rc = qsd_op_begin(env, qsd, trans, qi, NULL);
 808         /* can't get EDQUOT when reducing usage */
 809         if (rc == -EINPROGRESS)
 810                 rc = 0;
 811         return rc;
 812 }
 813
 814 static int osd_declare_attr_set(const struct lu_env *env,
 815                                 struct dt_object *dt,
 816                                 const struct lu_attr *attr,
 817                                 struct thandle *handle)
 818 {
 819         struct osd_thread_info  *info = osd_oti_get(env);
 820         struct osd_object       *obj = osd_dt_obj(dt);
 821         struct osd_device       *osd = osd_obj2dev(obj);
 822         dmu_tx_hold_t           *txh;
 823         struct osd_thandle      *oh;
 824         uint64_t                 bspace;
 825         uint32_t                 blksize;
 826         int                      rc = 0;
 827         bool                     found;
 828         ENTRY;
 829
 830
 831         LASSERT(handle != NULL);
 832         LASSERT(osd_invariant(obj));
 833
 834         oh = container_of0(handle, struct osd_thandle, ot_super);
 835
 836         down_read(&obj->oo_guard);
 837         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 838                 GOTO(out, rc = 0);
 839
 840         LASSERT(obj->oo_sa_hdl != NULL);
 841         LASSERT(oh->ot_tx != NULL);
 842         /* regular attributes are part of the bonus buffer */
 843         /* let's check whether this object is already part of
 844          * transaction.. */
 845         found = false;
 846         for (txh = list_head(&oh->ot_tx->tx_holds); txh;
 847              txh = list_next(&oh->ot_tx->tx_holds, txh)) {
 848                 if (txh->txh_dnode == NULL)
 849                         continue;
 850                 if (txh->txh_dnode->dn_object != obj->oo_dn->dn_object)
 851                         continue;
 852                 /* this object is part of the transaction already
 853                  * we don't need to declare bonus again */
 854                 found = true;
 855                 break;
 856         }
 857         if (!found)
 858                 dmu_tx_hold_bonus(oh->ot_tx, obj->oo_dn->dn_object);
 859         if (oh->ot_tx->tx_err != 0)
 860                 GOTO(out, rc = -oh->ot_tx->tx_err);
 861
 862         if (attr && attr->la_valid & LA_FLAGS) {
 863                 /* LMA is usually a part of bonus, no need to declare
 864                  * anything else */
 865         }
 866
 867         if (attr && (attr->la_valid & (LA_UID | LA_GID))) {
 868                 sa_object_size(obj->oo_sa_hdl, &blksize, &bspace);
 869                 bspace = toqb(bspace * blksize);
 870         }
 871
 872         if (attr && attr->la_valid & LA_UID) {
 873                 /* quota enforcement for user */
 874                 if (attr->la_uid != obj->oo_attr.la_uid) {
 875                         rc = qsd_transfer(env, osd->od_quota_slave,
 876                                           &oh->ot_quota_trans, USRQUOTA,
 877                                           obj->oo_attr.la_uid, attr->la_uid,
 878                                           bspace, &info->oti_qi);
 879                         if (rc)
 880                                 GOTO(out, rc);
 881                 }
 882         }
 883         if (attr && attr->la_valid & LA_GID) {
 884                 /* quota enforcement for group */
 885                 if (attr->la_gid != obj->oo_attr.la_gid) {
 886                         rc = qsd_transfer(env, osd->od_quota_slave,
 887                                           &oh->ot_quota_trans, GRPQUOTA,
 888                                           obj->oo_attr.la_gid, attr->la_gid,
 889                                           bspace, &info->oti_qi);
 890                         if (rc)
 891                                 GOTO(out, rc);
 892                 }
 893         }
 894
 895 out:
 896         up_read(&obj->oo_guard);
 897         RETURN(rc);
 898 }
 899
 900 /*
 901  * Set the attributes of an object
 902  *
 903  * The transaction passed to this routine must have
 904  * dmu_tx_hold_bonus(tx, oid) called and then assigned
 905  * to a transaction group.
 906  */
 907 static int osd_attr_set(const struct lu_env *env, struct dt_object *dt,
 908                         const struct lu_attr *la, struct thandle *handle)
 909 {
 910         struct osd_thread_info  *info = osd_oti_get(env);
 911         sa_bulk_attr_t          *bulk = osd_oti_get(env)->oti_attr_bulk;
 912         struct osd_object       *obj = osd_dt_obj(dt);
 913         struct osd_device       *osd = osd_obj2dev(obj);
 914         struct osd_thandle      *oh;
 915         struct osa_attr         *osa = &info->oti_osa;
 916         __u64                    valid = la->la_valid;
 917         int                      cnt;
 918         int                      rc = 0;
 919
 920         ENTRY;
 921
 922         down_read(&obj->oo_guard);
 923         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
 924                 GOTO(out, rc = -ENOENT);
 925
 926         LASSERT(handle != NULL);
 927         LASSERT(osd_invariant(obj));
 928         LASSERT(obj->oo_sa_hdl);
 929
 930         oh = container_of0(handle, struct osd_thandle, ot_super);
 931         /* Assert that the transaction has been assigned to a
 932            transaction group. */
 933         LASSERT(oh->ot_tx->tx_txg != 0);
 934
 935         /* Only allow set size for regular file */
 936         if (!S_ISREG(dt->do_lu.lo_header->loh_attr))
 937                 valid &= ~(LA_SIZE | LA_BLOCKS);
 938
 939         if (valid & LA_CTIME && la->la_ctime == obj->oo_attr.la_ctime)
 940                 valid &= ~LA_CTIME;
 941
 942         if (valid & LA_MTIME && la->la_mtime == obj->oo_attr.la_mtime)
 943                 valid &= ~LA_MTIME;
 944
 945         if (valid & LA_ATIME && la->la_atime == obj->oo_attr.la_atime)
 946                 valid &= ~LA_ATIME;
 947
 948         if (valid == 0)
 949                 GOTO(out, rc = 0);
 950
 951         if (valid & LA_FLAGS) {
 952                 struct lustre_mdt_attrs *lma;
 953                 struct lu_buf buf;
 954
 955                 if (la->la_flags & LUSTRE_LMA_FL_MASKS) {
 956                         CLASSERT(sizeof(info->oti_buf) >= sizeof(*lma));
 957                         lma = (struct lustre_mdt_attrs *)&info->oti_buf;
 958                         buf.lb_buf = lma;
 959                         buf.lb_len = sizeof(info->oti_buf);
 960                         rc = osd_xattr_get(env, &obj->oo_dt, &buf,
 961                                            XATTR_NAME_LMA);
 962                         if (rc > 0) {
 963                                 lma->lma_incompat =
 964                                         le32_to_cpu(lma->lma_incompat);
 965                                 lma->lma_incompat |=
 966                                         lustre_to_lma_flags(la->la_flags);
 967                                 lma->lma_incompat =
 968                                         cpu_to_le32(lma->lma_incompat);
 969                                 buf.lb_buf = lma;
 970                                 buf.lb_len = sizeof(*lma);
 971                                 rc = osd_xattr_set_internal(env, obj, &buf,
 972                                                             XATTR_NAME_LMA,
 973                                                             LU_XATTR_REPLACE,
 974                                                             oh);
 975                         }
 976                         if (rc < 0) {
 977                                 CWARN("%s: failed to set LMA flags: rc = %d\n",
 978                                        osd->od_svname, rc);
 979                                 RETURN(rc);
 980                         }
 981                 }
 982         }
 983
 984         write_lock(&obj->oo_attr_lock);
 985         cnt = 0;
 986         if (valid & LA_ATIME) {
 987                 osa->atime[0] = obj->oo_attr.la_atime = la->la_atime;
 988                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(osd), NULL,
 989                                  osa->atime, 16);
 990         }
 991         if (valid & LA_MTIME) {
 992                 osa->mtime[0] = obj->oo_attr.la_mtime = la->la_mtime;
 993                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(osd), NULL,
 994                                  osa->mtime, 16);
 995         }
 996         if (valid & LA_CTIME) {
 997                 osa->ctime[0] = obj->oo_attr.la_ctime = la->la_ctime;
 998                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CTIME(osd), NULL,
 999                                  osa->ctime, 16);
1000         }
1001         if (valid & LA_MODE) {
1002                 /* mode is stored along with type, so read it first */
1003                 obj->oo_attr.la_mode = (obj->oo_attr.la_mode & S_IFMT) |
1004                         (la->la_mode & ~S_IFMT);
1005                 osa->mode = obj->oo_attr.la_mode;
1006                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(osd), NULL,
1007                                  &osa->mode, 8);
1008         }
1009         if (valid & LA_SIZE) {
1010                 osa->size = obj->oo_attr.la_size = la->la_size;
1011                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_SIZE(osd), NULL,
1012                                  &osa->size, 8);
1013         }
1014         if (valid & LA_NLINK) {
1015                 osa->nlink = obj->oo_attr.la_nlink = la->la_nlink;
1016                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_LINKS(osd), NULL,
1017                                  &osa->nlink, 8);
1018         }
1019         if (valid & LA_RDEV) {
1020                 osa->rdev = obj->oo_attr.la_rdev = la->la_rdev;
1021                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_RDEV(osd), NULL,
1022                                  &osa->rdev, 8);
1023         }
1024         if (valid & LA_FLAGS) {
1025                 osa->flags = attrs_fs2zfs(la->la_flags);
1026                 /* many flags are not supported by zfs, so ensure a good cached
1027                  * copy */
1028                 obj->oo_attr.la_flags = attrs_zfs2fs(osa->flags);
1029                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(osd), NULL,
1030                                  &osa->flags, 8);
1031         }
1032         if (valid & LA_UID) {
1033                 osa->uid = obj->oo_attr.la_uid = la->la_uid;
1034                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_UID(osd), NULL,
1035                                  &osa->uid, 8);
1036         }
1037         if (valid & LA_GID) {
1038                 osa->gid = obj->oo_attr.la_gid = la->la_gid;
1039                 SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GID(osd), NULL,
1040                                  &osa->gid, 8);
1041         }
1042         obj->oo_attr.la_valid |= valid;
1043         write_unlock(&obj->oo_attr_lock);
1044
1045         LASSERT(cnt <= ARRAY_SIZE(osd_oti_get(env)->oti_attr_bulk));
1046         rc = osd_object_sa_bulk_update(obj, bulk, cnt, oh);
1047
1048 out:
1049         up_read(&obj->oo_guard);
1050         RETURN(rc);
1051 }
1052
1053 /*
1054  * Object creation.
1055  *
1056  * XXX temporary solution.
1057  */
1058
1059 static void osd_ah_init(const struct lu_env *env, struct dt_allocation_hint *ah,
1060                         struct dt_object *parent, struct dt_object *child,
1061                         umode_t child_mode)
1062 {
1063         LASSERT(ah);
1064
1065         ah->dah_parent = parent;
1066         ah->dah_mode = child_mode;
1067
1068         if (parent != NULL && !dt_object_remote(parent)) {
1069                 /* will help to find FID->ino at dt_insert("..") */
1070                 struct osd_object *pobj = osd_dt_obj(parent);
1071
1072                 osd_idc_find_and_init(env, osd_obj2dev(pobj), pobj);
1073         }
1074 }
1075
1076 static int osd_declare_object_create(const struct lu_env *env,
1077                                      struct dt_object *dt,
1078                                      struct lu_attr *attr,
1079                                      struct dt_allocation_hint *hint,
1080                                      struct dt_object_format *dof,
1081                                      struct thandle *handle)
1082 {
1083         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
1084         struct osd_object       *obj = osd_dt_obj(dt);
1085         struct osd_device       *osd = osd_obj2dev(obj);
1086         struct osd_thandle      *oh;
1087         uint64_t                 zapid;
1088         int                      rc, dnode_size;
1089         ENTRY;
1090
1091         LASSERT(dof);
1092
1093         switch (dof->dof_type) {
1094                 case DFT_REGULAR:
1095                 case DFT_SYM:
1096                 case DFT_NODE:
1097                         if (obj->oo_dt.do_body_ops == NULL)
1098                                 obj->oo_dt.do_body_ops = &osd_body_ops;
1099                         break;
1100                 default:
1101                         break;
1102         }
1103
1104         LASSERT(handle != NULL);
1105         oh = container_of0(handle, struct osd_thandle, ot_super);
1106         LASSERT(oh->ot_tx != NULL);
1107
1108         /* this is the minimum set of EAs on every Lustre object */
1109         obj->oo_ea_in_bonus = ZFS_SA_BASE_ATTR_SIZE +
1110                                 sizeof(__u64) + /* VBR VERSION */
1111                                 sizeof(struct lustre_mdt_attrs); /* LMA */
1112         /* reserve 32 bytes for extra stuff like ACLs */
1113         dnode_size = size_roundup_power2(obj->oo_ea_in_bonus + 32);
1114
1115         switch (dof->dof_type) {
1116                 case DFT_DIR:
1117                         dt->do_index_ops = &osd_dir_ops;
1118                 case DFT_INDEX:
1119                         /* for zap create */
1120                         dmu_tx_hold_zap(oh->ot_tx, DMU_NEW_OBJECT, FALSE, NULL);
1121                         dmu_tx_hold_sa_create(oh->ot_tx, dnode_size);
1122                         break;
1123                 case DFT_REGULAR:
1124                 case DFT_SYM:
1125                 case DFT_NODE:
1126                         /* first, we'll create new object */
1127                         dmu_tx_hold_sa_create(oh->ot_tx, dnode_size);
1128                         break;
1129
1130                 default:
1131                         LBUG();
1132                         break;
1133         }
1134
1135         /* and we'll add it to some mapping */
1136         zapid = osd_get_name_n_idx(env, osd, fid, NULL, 0);
1137         dmu_tx_hold_zap(oh->ot_tx, zapid, TRUE, NULL);
1138
1139         /* will help to find FID->ino mapping at dt_insert() */
1140         osd_idc_find_and_init(env, osd, obj);
1141
1142         rc = osd_declare_quota(env, osd, attr->la_uid, attr->la_gid, 1, oh,
1143                                false, NULL, false);
1144
1145         RETURN(rc);
1146 }
1147
1148 int __osd_attr_init(const struct lu_env *env, struct osd_device *osd,
1149                     sa_handle_t *sa_hdl, dmu_tx_t *tx,
1150                     struct lu_attr *la, uint64_t parent)
1151 {
1152         sa_bulk_attr_t  *bulk = osd_oti_get(env)->oti_attr_bulk;
1153         struct osa_attr *osa = &osd_oti_get(env)->oti_osa;
1154         uint64_t         gen;
1155         uint64_t         crtime[2];
1156         timestruc_t      now;
1157         int              cnt;
1158         int              rc;
1159
1160         LASSERT(sa_hdl);
1161
1162         gen = dmu_tx_get_txg(tx);
1163         gethrestime(&now);
1164         ZFS_TIME_ENCODE(&now, crtime);
1165
1166         osa->atime[0] = la->la_atime;
1167         osa->ctime[0] = la->la_ctime;
1168         osa->mtime[0] = la->la_mtime;
1169         osa->mode = la->la_mode;
1170         osa->uid = la->la_uid;
1171         osa->gid = la->la_gid;
1172         osa->rdev = la->la_rdev;
1173         osa->nlink = la->la_nlink;
1174         osa->flags = attrs_fs2zfs(la->la_flags);
1175         osa->size  = la->la_size;
1176
1177         /*
1178          * we need to create all SA below upon object create.
1179          *
1180          * XXX The attribute order matters since the accounting callback relies
1181          * on static offsets (i.e. SA_*_OFFSET, see zfs_space_delta_cb()) to
1182          * look up the UID/GID attributes. Moreover, the callback does not seem
1183          * to support the spill block.
1184          * We define attributes in the same order as SA_*_OFFSET in order to
1185          * work around the problem. See ORI-610.
1186          */
1187         cnt = 0;
1188         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(osd), NULL, &osa->mode, 8);
1189         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_SIZE(osd), NULL, &osa->size, 8);
1190         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GEN(osd), NULL, &gen, 8);
1191         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_UID(osd), NULL, &osa->uid, 8);
1192         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_GID(osd), NULL, &osa->gid, 8);
1193         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_PARENT(osd), NULL, &parent, 8);
1194         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_FLAGS(osd), NULL, &osa->flags, 8);
1195         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(osd), NULL, osa->atime, 16);
1196         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(osd), NULL, osa->mtime, 16);
1197         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CTIME(osd), NULL, osa->ctime, 16);
1198         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_CRTIME(osd), NULL, crtime, 16);
1199         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_LINKS(osd), NULL, &osa->nlink, 8);
1200         SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_RDEV(osd), NULL, &osa->rdev, 8);
1201         LASSERT(cnt <= ARRAY_SIZE(osd_oti_get(env)->oti_attr_bulk));
1202
1203         rc = -sa_replace_all_by_template(sa_hdl, bulk, cnt, tx);
1204
1205         return rc;
1206 }
1207
1208 static int osd_find_new_dnode(const struct lu_env *env, dmu_tx_t *tx,
1209                               uint64_t oid, dnode_t **dnp)
1210 {
1211         dmu_tx_hold_t *txh;
1212         int rc = 0;
1213
1214         /* take dnode_t from tx to save on dnode#->dnode_t lookup */
1215         for (txh = list_tail(&tx->tx_holds); txh;
1216              txh = list_prev(&tx->tx_holds, txh)) {
1217                 dnode_t *dn = txh->txh_dnode;
1218                 dmu_buf_impl_t *db;
1219
1220                 if (dn == NULL)
1221                         continue;
1222                 if (dn->dn_object != oid)
1223                         continue;
1224                 db = dn->dn_bonus;
1225                 if (db == NULL) {
1226                         rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1227                         if (dn->dn_bonus == NULL)
1228                                 dbuf_create_bonus(dn);
1229                         rw_exit(&dn->dn_struct_rwlock);
1230                 }
1231                 db = dn->dn_bonus;
1232                 LASSERT(db);
1233                 LASSERT(dn->dn_handle);
1234                 DB_DNODE_ENTER(db);
1235                 if (refcount_add(&db->db_holds, osd_obj_tag) == 1) {
1236                         refcount_add(&dn->dn_holds, tag);
1237                         atomic_inc_32(&dn->dn_dbufs_count);
1238                 }
1239                 *dnp = dn;
1240                 dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH);
1241                 break;
1242         }
1243
1244         if (unlikely(*dnp == NULL))
1245                 rc = __osd_obj2dnode(env, tx->tx_objset, oid, dnp);
1246
1247         return rc;
1248 }
1249
1250 /*
1251  * The transaction passed to this routine must have
1252  * dmu_tx_hold_bonus(tx, DMU_NEW_OBJECT) called and then assigned
1253  * to a transaction group.
1254  */
1255 int __osd_object_create(const struct lu_env *env, struct osd_object *obj,
1256                         dnode_t **dnp, dmu_tx_t *tx, struct lu_attr *la)
1257 {
1258         struct osd_device   *osd = osd_obj2dev(obj);
1259         const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
1260         dmu_object_type_t    type = DMU_OT_PLAIN_FILE_CONTENTS;
1261         uint64_t oid;
1262
1263         /* Use DMU_OTN_UINT8_METADATA for local objects so their data blocks
1264          * would get an additional ditto copy */
1265         if (unlikely(S_ISREG(la->la_mode) &&
1266                      fid_seq_is_local_file(fid_seq(fid))))
1267                 type = DMU_OTN_UINT8_METADATA;
1268
1269         /* Create a new DMU object using the default dnode size. */
1270         oid = osd_dmu_object_alloc(osd->od_os, type, 0, 0, tx);
1271
1272         LASSERT(la->la_valid & LA_MODE);
1273         la->la_size = 0;
1274         la->la_nlink = 1;
1275
1276         return osd_find_new_dnode(env, tx, oid, dnp);
1277 }
1278
1279 /*
1280  * The transaction passed to this routine must have
1281  * dmu_tx_hold_zap(tx, DMU_NEW_OBJECT, ...) called and then assigned
1282  * to a transaction group.
1283  *
1284  * Using ZAP_FLAG_HASH64 will force the ZAP to always be a FAT ZAP.
1285  * This is fine for directories today, because storing the FID in the dirent
1286  * will also require a FAT ZAP.  If there is a new type of micro ZAP created
1287  * then we might need to re-evaluate the use of this flag and instead do
1288  * a conversion from the different internal ZAP hash formats being used. */
1289 int __osd_zap_create(const struct lu_env *env, struct osd_device *osd,
1290                      dnode_t **dnp, dmu_tx_t *tx, struct lu_attr *la,
1291                      zap_flags_t flags)
1292 {
1293         uint64_t oid;
1294
1295         /* Assert that the transaction has been assigned to a
1296            transaction group. */
1297         LASSERT(tx->tx_txg != 0);
1298         *dnp = NULL;
1299
1300         oid = osd_zap_create_flags(osd->od_os, 0, flags | ZAP_FLAG_HASH64,
1301                                    DMU_OT_DIRECTORY_CONTENTS,
1302                                    14, /* == ZFS fzap_default_blockshift */
1303                                    DN_MAX_INDBLKSHIFT, /* indirect blockshift */
1304                                    0, tx);
1305
1306         la->la_size = 2;
1307         la->la_nlink = 1;
1308
1309         return osd_find_new_dnode(env, tx, oid, dnp);
1310 }
1311
1312 static dnode_t *osd_mkidx(const struct lu_env *env, struct osd_object *obj,
1313                           struct lu_attr *la, struct osd_thandle *oh)
1314 {
1315         dnode_t *dn;
1316         int rc;
1317
1318         /* Index file should be created as regular file in order not to confuse
1319          * ZPL which could interpret them as directory.
1320          * We set ZAP_FLAG_UINT64_KEY to let ZFS know than we are going to use
1321          * binary keys */
1322         LASSERT(S_ISREG(la->la_mode));
1323         rc = __osd_zap_create(env, osd_obj2dev(obj), &dn, oh->ot_tx, la,
1324                               ZAP_FLAG_UINT64_KEY);
1325         if (rc)
1326                 return ERR_PTR(rc);
1327         return dn;
1328 }
1329
1330 static dnode_t *osd_mkdir(const struct lu_env *env, struct osd_object *obj,
1331                           struct lu_attr *la, struct osd_thandle *oh)
1332 {
1333         dnode_t *dn;
1334         int rc;
1335
1336         LASSERT(S_ISDIR(la->la_mode));
1337         rc = __osd_zap_create(env, osd_obj2dev(obj), &dn, oh->ot_tx, la, 0);
1338         if (rc)
1339                 return ERR_PTR(rc);
1340         return dn;
1341 }
1342
1343 static dnode_t *osd_mkreg(const struct lu_env *env, struct osd_object *obj,
1344                           struct lu_attr *la, struct osd_thandle *oh)
1345 {
1346         const struct lu_fid *fid = lu_object_fid(&obj->oo_dt.do_lu);
1347         struct osd_device *osd = osd_obj2dev(obj);
1348         dnode_t *dn;
1349         int rc;
1350
1351         LASSERT(S_ISREG(la->la_mode));
1352         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1353         if (rc)
1354                 return ERR_PTR(rc);
1355
1356         if ((fid_is_idif(fid) || fid_is_norm(fid) || fid_is_echo(fid)) &&
1357             osd->od_is_ost) {
1358                 /* The minimum block size must be at least page size otherwise
1359                  * it will break the assumption in tgt_thread_big_cache where
1360                  * the array size is PTLRPC_MAX_BRW_PAGES. It will also affect
1361                  * RDMA due to subpage transfer size */
1362                 rc = -dmu_object_set_blocksize(osd->od_os, dn->dn_object,
1363                                                PAGE_SIZE, 0, oh->ot_tx);
1364                 if (unlikely(rc)) {
1365                         CERROR("%s: can't change blocksize: %d\n",
1366                                osd->od_svname, rc);
1367                         return ERR_PTR(rc);
1368                 }
1369         }
1370
1371         return dn;
1372 }
1373
1374 static dnode_t *osd_mksym(const struct lu_env *env, struct osd_object *obj,
1375                           struct lu_attr *la, struct osd_thandle *oh)
1376 {
1377         dnode_t *dn;
1378         int rc;
1379
1380         LASSERT(S_ISLNK(la->la_mode));
1381         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1382         if (rc)
1383                 return ERR_PTR(rc);
1384         return dn;
1385 }
1386
1387 static dnode_t *osd_mknod(const struct lu_env *env, struct osd_object *obj,
1388                           struct lu_attr *la, struct osd_thandle *oh)
1389 {
1390         dnode_t *dn;
1391         int rc;
1392
1393         if (S_ISCHR(la->la_mode) || S_ISBLK(la->la_mode))
1394                 la->la_valid |= LA_RDEV;
1395
1396         rc = __osd_object_create(env, obj, &dn, oh->ot_tx, la);
1397         if (rc)
1398                 return ERR_PTR(rc);
1399         return dn;
1400 }
1401
1402 typedef dnode_t *(*osd_obj_type_f)(const struct lu_env *env,
1403                                    struct osd_object *obj,
1404                                    struct lu_attr *la,
1405                                    struct osd_thandle *oh);
1406
1407 static osd_obj_type_f osd_create_type_f(enum dt_format_type type)
1408 {
1409         osd_obj_type_f result;
1410
1411         switch (type) {
1412         case DFT_DIR:
1413                 result = osd_mkdir;
1414                 break;
1415         case DFT_INDEX:
1416                 result = osd_mkidx;
1417                 break;
1418         case DFT_REGULAR:
1419                 result = osd_mkreg;
1420                 break;
1421         case DFT_SYM:
1422                 result = osd_mksym;
1423                 break;
1424         case DFT_NODE:
1425                 result = osd_mknod;
1426                 break;
1427         default:
1428                 LBUG();
1429                 break;
1430         }
1431         return result;
1432 }
1433
1434 /*
1435  * Concurrency: @dt is write locked.
1436  */
1437 static int osd_object_create(const struct lu_env *env, struct dt_object *dt,
1438                              struct lu_attr *attr,
1439                              struct dt_allocation_hint *hint,
1440                              struct dt_object_format *dof,
1441                              struct thandle *th)
1442 {
1443         struct osd_thread_info  *info = osd_oti_get(env);
1444         struct lustre_mdt_attrs *lma = &info->oti_mdt_attrs;
1445         struct zpl_direntry     *zde = &info->oti_zde.lzd_reg;
1446         const struct lu_fid     *fid = lu_object_fid(&dt->do_lu);
1447         struct osd_object       *obj = osd_dt_obj(dt);
1448         struct osd_device       *osd = osd_obj2dev(obj);
1449         char                    *buf = info->oti_str;
1450         struct osd_thandle      *oh;
1451         dnode_t *dn = NULL;
1452         uint64_t                 zapid, parent = 0;
1453         int                      rc;
1454
1455         ENTRY;
1456
1457         /* concurrent create declarations should not see
1458          * the object inconsistent (db, attr, etc).
1459          * in regular cases acquisition should be cheap */
1460         down_write(&obj->oo_guard);
1461
1462         if (unlikely(dt_object_exists(dt)))
1463                 GOTO(out, rc = -EEXIST);
1464
1465         LASSERT(osd_invariant(obj));
1466         LASSERT(dof != NULL);
1467
1468         LASSERT(th != NULL);
1469         oh = container_of0(th, struct osd_thandle, ot_super);
1470
1471         LASSERT(obj->oo_dn == NULL);
1472
1473         /* to follow ZFS on-disk format we need
1474          * to initialize parent dnode properly */
1475         if (hint != NULL && hint->dah_parent != NULL &&
1476             !dt_object_remote(hint->dah_parent))
1477                 parent = osd_dt_obj(hint->dah_parent)->oo_dn->dn_object;
1478
1479         /* we may fix some attributes, better do not change the source */
1480         obj->oo_attr = *attr;
1481         obj->oo_attr.la_valid |= LA_SIZE | LA_NLINK | LA_TYPE;
1482
1483         dn = osd_create_type_f(dof->dof_type)(env, obj, &obj->oo_attr, oh);
1484         if (IS_ERR(dn)) {
1485                 rc = PTR_ERR(dn);
1486                 dn = NULL;
1487                 GOTO(out, rc);
1488         }
1489
1490         zde->zde_pad = 0;
1491         zde->zde_dnode = dn->dn_object;
1492         zde->zde_type = IFTODT(attr->la_mode & S_IFMT);
1493
1494         zapid = osd_get_name_n_idx(env, osd, fid, buf, sizeof(info->oti_str));
1495
1496         rc = -zap_add(osd->od_os, zapid, buf, 8, 1, zde, oh->ot_tx);
1497         if (rc)
1498                 GOTO(out, rc);
1499         obj->oo_dn = dn;
1500         /* Now add in all of the "SA" attributes */
1501         rc = osd_sa_handle_get(obj);
1502         if (rc)
1503                 GOTO(out, rc);
1504
1505         /* configure new osd object */
1506         parent = parent != 0 ? parent : zapid;
1507         rc = __osd_attr_init(env, osd, obj->oo_sa_hdl, oh->ot_tx,
1508                              &obj->oo_attr, parent);
1509         if (rc)
1510                 GOTO(out, rc);
1511
1512         /* XXX: oo_lma_flags */
1513         obj->oo_dt.do_lu.lo_header->loh_attr |= obj->oo_attr.la_mode & S_IFMT;
1514         if (likely(!fid_is_acct(lu_object_fid(&obj->oo_dt.do_lu))))
1515                 /* no body operations for accounting objects */
1516                 obj->oo_dt.do_body_ops = &osd_body_ops;
1517
1518         rc = -nvlist_alloc(&obj->oo_sa_xattr, NV_UNIQUE_NAME, KM_SLEEP);
1519         if (rc)
1520                 GOTO(out, rc);
1521
1522         /* initialize LMA */
1523         lustre_lma_init(lma, lu_object_fid(&obj->oo_dt.do_lu), 0, 0);
1524         lustre_lma_swab(lma);
1525         rc = -nvlist_add_byte_array(obj->oo_sa_xattr, XATTR_NAME_LMA,
1526                                     (uchar_t *)lma, sizeof(*lma));
1527         if (rc)
1528                 GOTO(out, rc);
1529         rc = __osd_sa_xattr_update(env, obj, oh);
1530         if (rc)
1531                 GOTO(out, rc);
1532         osd_idc_find_and_init(env, osd, obj);
1533
1534 out:
1535         if (unlikely(rc && dn)) {
1536                 dmu_object_free(osd->od_os, dn->dn_object, oh->ot_tx);
1537                 osd_dnode_rele(dn);
1538                 obj->oo_dn = NULL;
1539         } else if (!rc) {
1540                 obj->oo_dt.do_lu.lo_header->loh_attr |= LOHA_EXISTS;
1541         }
1542         up_write(&obj->oo_guard);
1543         RETURN(rc);
1544 }
1545
1546 static int osd_declare_object_ref_add(const struct lu_env *env,
1547                                       struct dt_object *dt,
1548                                       struct thandle *th)
1549 {
1550         return osd_declare_attr_set(env, dt, NULL, th);
1551 }
1552
1553 /*
1554  * Concurrency: @dt is write locked.
1555  */
1556 static int osd_object_ref_add(const struct lu_env *env,
1557                               struct dt_object *dt,
1558                               struct thandle *handle)
1559 {
1560         struct osd_object       *obj = osd_dt_obj(dt);
1561         struct osd_thandle      *oh;
1562         struct osd_device       *osd = osd_obj2dev(obj);
1563         uint64_t                 nlink;
1564         int rc;
1565
1566         ENTRY;
1567
1568         down_read(&obj->oo_guard);
1569         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
1570                 GOTO(out, rc = -ENOENT);
1571
1572         LASSERT(osd_invariant(obj));
1573         LASSERT(obj->oo_sa_hdl != NULL);
1574
1575         oh = container_of0(handle, struct osd_thandle, ot_super);
1576
1577         write_lock(&obj->oo_attr_lock);
1578         nlink = ++obj->oo_attr.la_nlink;
1579         write_unlock(&obj->oo_attr_lock);
1580
1581         rc = osd_object_sa_update(obj, SA_ZPL_LINKS(osd), &nlink, 8, oh);
1582
1583 out:
1584         up_read(&obj->oo_guard);
1585         RETURN(rc);
1586 }
1587
1588 static int osd_declare_object_ref_del(const struct lu_env *env,
1589                                       struct dt_object *dt,
1590                                       struct thandle *handle)
1591 {
1592         return osd_declare_attr_set(env, dt, NULL, handle);
1593 }
1594
1595 /*
1596  * Concurrency: @dt is write locked.
1597  */
1598 static int osd_object_ref_del(const struct lu_env *env,
1599                               struct dt_object *dt,
1600                               struct thandle *handle)
1601 {
1602         struct osd_object       *obj = osd_dt_obj(dt);
1603         struct osd_thandle      *oh;
1604         struct osd_device       *osd = osd_obj2dev(obj);
1605         uint64_t                 nlink;
1606         int                      rc;
1607
1608         ENTRY;
1609
1610         down_read(&obj->oo_guard);
1611
1612         if (unlikely(!dt_object_exists(dt) || obj->oo_destroyed))
1613                 GOTO(out, rc = -ENOENT);
1614
1615         LASSERT(osd_invariant(obj));
1616         LASSERT(obj->oo_sa_hdl != NULL);
1617
1618         oh = container_of0(handle, struct osd_thandle, ot_super);
1619         LASSERT(!lu_object_is_dying(dt->do_lu.lo_header));
1620
1621         write_lock(&obj->oo_attr_lock);
1622         nlink = --obj->oo_attr.la_nlink;
1623         write_unlock(&obj->oo_attr_lock);
1624
1625         rc = osd_object_sa_update(obj, SA_ZPL_LINKS(osd), &nlink, 8, oh);
1626
1627 out:
1628         up_read(&obj->oo_guard);
1629         RETURN(rc);
1630 }
1631
1632 static int osd_object_sync(const struct lu_env *env, struct dt_object *dt,
1633                            __u64 start, __u64 end)
1634 {
1635         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1636         ENTRY;
1637
1638         /* XXX: no other option than syncing the whole filesystem until we
1639          * support ZIL.  If the object tracked the txg that it was last
1640          * modified in, it could pass that txg here instead of "0".  Maybe
1641          * the changes are already committed, so no wait is needed at all? */
1642         if (!osd->od_dt_dev.dd_rdonly)
1643                 txg_wait_synced(dmu_objset_pool(osd->od_os), 0ULL);
1644
1645         RETURN(0);
1646 }
1647
1648 static int osd_invalidate(const struct lu_env *env, struct dt_object *dt)
1649 {
1650         return 0;
1651 }
1652
1653 static struct dt_object_operations osd_obj_ops = {
1654         .do_read_lock           = osd_object_read_lock,
1655         .do_write_lock          = osd_object_write_lock,
1656         .do_read_unlock         = osd_object_read_unlock,
1657         .do_write_unlock        = osd_object_write_unlock,
1658         .do_write_locked        = osd_object_write_locked,
1659         .do_attr_get            = osd_attr_get,
1660         .do_declare_attr_set    = osd_declare_attr_set,
1661         .do_attr_set            = osd_attr_set,
1662         .do_ah_init             = osd_ah_init,
1663         .do_declare_create      = osd_declare_object_create,
1664         .do_create              = osd_object_create,
1665         .do_declare_destroy     = osd_declare_object_destroy,
1666         .do_destroy             = osd_object_destroy,
1667         .do_index_try           = osd_index_try,
1668         .do_declare_ref_add     = osd_declare_object_ref_add,
1669         .do_ref_add             = osd_object_ref_add,
1670         .do_declare_ref_del     = osd_declare_object_ref_del,
1671         .do_ref_del             = osd_object_ref_del,
1672         .do_xattr_get           = osd_xattr_get,
1673         .do_declare_xattr_set   = osd_declare_xattr_set,
1674         .do_xattr_set           = osd_xattr_set,
1675         .do_declare_xattr_del   = osd_declare_xattr_del,
1676         .do_xattr_del           = osd_xattr_del,
1677         .do_xattr_list          = osd_xattr_list,
1678         .do_object_sync         = osd_object_sync,
1679         .do_invalidate          = osd_invalidate,
1680 };
1681
1682 static struct lu_object_operations osd_lu_obj_ops = {
1683         .loo_object_init        = osd_object_init,
1684         .loo_object_delete      = osd_object_delete,
1685         .loo_object_release     = osd_object_release,
1686         .loo_object_free        = osd_object_free,
1687         .loo_object_print       = osd_object_print,
1688         .loo_object_invariant   = osd_object_invariant,
1689 };
1690
1691 static int osd_otable_it_attr_get(const struct lu_env *env,
1692                                 struct dt_object *dt,
1693                                 struct lu_attr *attr)
1694 {
1695         attr->la_valid = 0;
1696         return 0;
1697 }
1698
1699 static struct dt_object_operations osd_obj_otable_it_ops = {
1700         .do_attr_get    = osd_otable_it_attr_get,
1701         .do_index_try   = osd_index_try,
1702 };